snapshot

name	snapshot
description	Creates a full dated backup of all project assets NOT checked into git. Captures annotations, labels, images, previews, manifests, configs, experiment docs, scrape reports, model weights, and skills — everything needed to restore the full project state and continue work. Use when the user says "snapshot", "backup", "archive", or "save state".

Project Snapshot — Full Asset Backup

What Gets Backed Up

Everything NOT in git that's needed to continue the project:

Category	Path	Why
Annotations	data/tmp/annotations/*.json	Hours of Sonnet labeling work
YOLO labels	data/tmp/labels/*.txt	Derived from annotations
Raw images	data/tmp/*.jpg, data/tmp/*.jpeg	Scraped images, irreplaceable
Preview QC images	data/tmp/preview/*.jpg	Visual verification evidence
Manifests	data/tmp/labeling_manifest.csv, data/tmp/manifest.json	Pipeline state
Pipeline events	data/tmp/pipeline_events.jsonl	Full audit trail
Canonical test set	configs/canonical_test_v2.json	Verified test config
Experiment docs	docs/experiments/*.md	Research history
Scrape reports	docs/scrape_reports/*.md	Data provenance
Research docs	docs/research/*.md	External references
Model weights	runs/	Trained model checkpoints
Freeze-split baseline	/tmp/finder_baseline/	Dataset split anchor
Skills	.claude/skills/*/	Agent instructions
Negatives	data/negatives/	Background images

How to Run

DATE=$(date +%Y%m%dT%H%M%S)
BACKUP_DIR="backups/full_${DATE}"
mkdir -p "$BACKUP_DIR"

# 1. Annotations + labels + manifests (small, critical)
tar czf "$BACKUP_DIR/annotations.tar.gz" \
    data/tmp/annotations/ \
    data/tmp/labels/ \
    data/tmp/labeling_manifest.csv \
    data/tmp/manifest.json \
    data/tmp/pipeline_events.jsonl \
    2>/dev/null

# 2. Raw images (large, irreplaceable)
tar czf "$BACKUP_DIR/images.tar.gz" \
    data/tmp/*.jpg data/tmp/*.jpeg \
    2>/dev/null

# 3. Preview QC images (large, regenerable but slow)
tar czf "$BACKUP_DIR/previews.tar.gz" \
    data/tmp/preview/ \
    2>/dev/null

# 4. Untracked docs + configs + skills
tar czf "$BACKUP_DIR/docs_and_config.tar.gz" \
    docs/experiments/ \
    docs/research/ \
    docs/scrape_reports/ \
    docs/model_cards/ \
    docs/ARCHITECTURE_REVIEW_v7.md \
    docs/LABELING_RULES_v7_DRAFT.md \
    configs/canonical_test_v2.json \
    .claude/skills/ml-researcher/ \
    .claude/skills/sign-crop-labeler/ \
    .claude/skills/snapshot/ \
    2>/dev/null

# 5. Model weights (regenerable but expensive)
tar czf "$BACKUP_DIR/model_weights.tar.gz" \
    runs/ \
    2>/dev/null

# 6. Negatives + ingest manifests
tar czf "$BACKUP_DIR/negatives_and_ingest.tar.gz" \
    data/negatives/ \
    data/ingest/*/scrape_manifest.json \
    2>/dev/null

# Summary
echo "=== Snapshot: $BACKUP_DIR ==="
ls -lh "$BACKUP_DIR/"
echo ""
du -sh "$BACKUP_DIR/"

How to Restore

BACKUP_DIR="backups/full_YYYYMMDDTHHMMSS"

# Restore everything
for f in "$BACKUP_DIR"/*.tar.gz; do
    echo "Restoring $f..."
    tar xzf "$f"
done

echo "Done. Run 'python scripts/build_finder_dataset.py --classes 0 --seed 42' to rebuild dataset."

When to Snapshot

• Before any destructive operation (relabeling, dataset rebuild, annotation cleanup)
• At the end of a session (preserve progress)
• Before changing labeling rules (v5→v6→v7 transitions)
• Before merging data from multiple sources